0% Markup Video APIs: The Real Cost per Ad Clip
Sume, Cloudflare and INFRO illustrate why provider-rate video pricing differs from finished-ad cost: funding fees, concurrency, retries and creative approval.
A team evaluating Sume should compare cost per approved ad, not just the advertised provider rate. A successful render, a usable creative and a credit-card payment belong in different columns of the budget.
Quick Answer
Sume publishes provider-rate video pricing with no subscription, a minimum top-up of $5 and credits valid for 365 days. Its pricing page also ties higher concurrency to recent funding. Those are meaningful commercial terms, but they do not specify how many generated candidates a team will need before approving an ad. Sume pricing

The same distinction applies elsewhere. Cloudflare says its Unified Billing passes through inference pricing without markup, but adds a 5% fee when credits are purchased. Zero inference markup therefore does not necessarily mean zero intermediary cost. Cloudflare Unified Billing
The useful comparison has three parts: the exact model tariff, the funding and throughput conditions, and the amount of paid generation that survives creative approval.
What the published specifications actually establish
The following is a comparison of commercial structures, not a ranking of output quality. Match the model, resolution, audio setting and duration before comparing rates.
| Route | Published generation price | Separate funding or platform condition |
|---|---|---|
| Sume | Veo 3.1 Fast: $0.08 per second at 720p with audio off | Top up $25 within 28 days for 10 concurrent jobs, $100 for 20, or $1,000 for 40. These are funding thresholds, not stated charges per clip. Sume model prices and concurrency terms |
| Cloudflare Unified Billing | Veo 3.1 Fast: $0.08 per second at 720p without audio. The same model is $0.10 with audio at that resolution. Cloudflare model rate card | Credit purchases carry a 5% fee, separate from the inference tariff. Cloudflare billing terms |
| INFRO with customer-owned provider accounts | The connected provider bills the customer at their own rates; this is not a single reseller model tariff | The Developer gateway plan is $35 per month, excluding tax, in addition to provider spend. INFRO pricing |
The matching Sume and Cloudflare entries establish a matching advertised rate for that configuration. They do not establish identical latency, available capacity or approval rates across the services. Nor does a matching entry validate an entire catalog's price-parity claim.
INFRO belongs in a different accounting category. Its video API routes through a customer's own provider accounts, rather than replacing those accounts with a prepaid reseller balance. It is relevant to the zero-markup comparison, but its subscription must be allocated separately from generation spend. INFRO video API documentation
Concurrency changes the funding requirement, not the tariff
Sume's funding-linked concurrency creates a practical distinction between buying capacity access and consuming credits. A team can need more money in the account to run a burst than the burst ultimately consumes. Do not book the entire top-up as the cost of that batch if a usable balance remains afterward.
Conversely, do not ignore the funding requirement when comparing services. A small campaign with a tight deadline may need parallel jobs even when its expected generation spend is modest. Confirm the applicable workspace cap before building the launch schedule, and account for the recent-top-up window rather than treating an unlocked tier as permanent.
Cloudflare documents a different quantity: Unified Billing permits 200 requests per 60 seconds per gateway using Cloudflare-managed credentials. That limit does not apply to BYOK requests. It is a request-rate allowance, not confirmation that the same number of video renders can run simultaneously. Cloudflare AI Gateway limits
For capacity planning, record both the submission limit and the number of accepted jobs still queued or running. Neither a requests-per-minute figure nor a per-second price establishes turnaround time.
Retries: distinguish a refused request from a rejected creative
Sume documents separate request-rate and workspace in-flight limits. A submission beyond the in-flight cap returns 429 queue_full, is refused rather than queued, and is not charged. Its instructions say to wait for capacity and retry that refusal with the same Idempotency-Key. An insufficient-balance refusal requires funding and a new key instead. Submitted jobs reserve balance; failed or canceled jobs release that reservation through a refund. Sume limits and billing
That creates three different cost events:
- Refused submission: no generation was accepted under the documented refusal conditions.
- Failed generation: follow the service's settlement and refund rules.
- Completed but unsuitable candidate: this is creative rejection, not automatically a technical failure entitled to a refund.
Cloudflare's request-handling documentation describes configurable retries and backoff for failed requests. That is a transport mechanism, not a promise that a completed video will satisfy an ad brief. Do not treat retry configuration as evidence of free creative revisions. Cloudflare request handling
The job lifecycle also matters. INFRO says fallback can try another connection when submission fails, but an accepted render is not moved to another connection if it later fails. Submission resilience and render recovery are distinct capabilities. INFRO asynchronous video lifecycle
Calculate the approved-clip cost, then add finishing
Use this accounting rule for a batch:
Cost per finished ad = (settled generation charges after refunds + attributable funding fees + finishing costs + allocated platform charges) ÷ approved final ads.
Keep top-ups in a separate cash-flow ledger. Otherwise, buying credits and later consuming them can be counted twice.
For an illustration, not an observed acceptance rate, assume three completed six-second candidates produce one approved clip. At Cloudflare's published 720p, audio-off rate of $0.08 per second, generation consumes $1.44 rather than the $0.48 required for a single candidate. Those duration and price settings are listed in its Veo model specifications.
Allocating the credit-purchase fee proportionally adds 5%, bringing that generation-and-funding subtotal to $1.512, or about $1.51. This follows Cloudflare's published fee, not a separate charge assessed on each render. The actual card payment depends on how much credit the team purchases. Cloudflare credit-purchase fee
That subtotal still excludes editing, voiceover, captions, assembly and human review where required. Measure those separately. If the approved ad uses several shots, include the paid candidates for every shot, not just the last selected render.
Trade-offs and who each structure fits
Sume fits a prepaid workflow when its published model options and funding-linked parallelism match the team's workload. The trade-off to examine is capital committed for burst capacity versus credits actually consumed. Its headline rate alone cannot answer that question.
Cloudflare fits teams that value consolidated funding and gateway controls. Its published structure makes the funding fee a separate, visible part of the comparison. Evaluate the exact billing path: using provider credentials is not the same arrangement as spending managed Unified Billing credits.
A BYOK gateway such as INFRO fits teams that want to retain their provider accounts and rates. Its fixed platform subscription must still be assigned somewhere in the production budget, even though model spend carries no gateway markup.
Before committing a campaign, run a representative batch with explicit approval criteria. Record accepted jobs, refunds, approved assets, elapsed completion time, remaining prepaid balance and finishing work. Choose the route that meets the deadline at an acceptable all-in cost per approved ad, not the one with the smallest isolated tariff.
FAQ
Does provider pricing guarantee the lowest finished-ad cost?
No. It describes an inference-pricing relationship. Funding fees, platform charges, finishing work and the number of paid candidates needed for approval determine the broader cost.
Is a larger top-up the same as a more expensive clip?
No. A top-up funds an account. Consumed credits become generation expense; remaining credits remain prepaid purchasing power, subject to the service's validity and refund terms. Track cash committed separately from batch consumption.
Does an API retry mean another paid generation?
Not necessarily. First establish whether the original submission was refused, accepted and still running, or completed. Follow the documented error-specific retry and idempotency instructions rather than blindly submitting a new job.
What should count as a finished ad in the cost calculation?
Define an approved final deliverable before the batch starts: required length, aspect ratio, audio, branding and review status. A generated candidate should enter the denominator only if it meets that definition; otherwise, its settled cost belongs in the numerator without adding another finished ad.